This paper proposes a novel method called Spatial Feature Transform (SFT) for image super-resolution (SR) to recover realistic textures. The method uses semantic segmentation probability maps to condition the SR network, enabling it to generate more realistic and visually pleasing textures. The SFT layer modulates features of intermediate layers in the network using affine transformation parameters derived from the segmentation maps. This allows the SR network to generate high-resolution images with just a single forward pass, conditioned on categorical priors. The method is parameter-efficient, can be easily integrated into existing SR networks, and is extensible to other types of priors. Experiments show that the proposed SFT-GAN outperforms state-of-the-art SRGAN and EnhanceNet in terms of perceptual quality and texture realism. The method is particularly effective in generating textures for different semantic regions, such as buildings, plants, and water. The SFT layer is capable of handling spatial information and provides fine-grained control over feature modulation. The method is also robust to out-of-category examples and can be applied to various types of images, including outdoor scenes. The results demonstrate that the SFT-GAN generates more realistic textures compared to other methods, making it a promising approach for image super-resolution.This paper proposes a novel method called Spatial Feature Transform (SFT) for image super-resolution (SR) to recover realistic textures. The method uses semantic segmentation probability maps to condition the SR network, enabling it to generate more realistic and visually pleasing textures. The SFT layer modulates features of intermediate layers in the network using affine transformation parameters derived from the segmentation maps. This allows the SR network to generate high-resolution images with just a single forward pass, conditioned on categorical priors. The method is parameter-efficient, can be easily integrated into existing SR networks, and is extensible to other types of priors. Experiments show that the proposed SFT-GAN outperforms state-of-the-art SRGAN and EnhanceNet in terms of perceptual quality and texture realism. The method is particularly effective in generating textures for different semantic regions, such as buildings, plants, and water. The SFT layer is capable of handling spatial information and provides fine-grained control over feature modulation. The method is also robust to out-of-category examples and can be applied to various types of images, including outdoor scenes. The results demonstrate that the SFT-GAN generates more realistic textures compared to other methods, making it a promising approach for image super-resolution.